Glandular Structure Segmentation from Colon Histopathology Images By : Ramin Nateghi

advertisement
Glandular Structure Segmentation from Colon
Histopathology Images
By : Ramin Nateghi
Shiraz University of Technology (SUTECH) Team
GlaS Contest , September 2015
1
ABOUT GLANDS
Typically a gland includes three main structures:
 Lumen – Cytoplasm - Cell nuclei
Cytoplasm
Lumen
Cell Nuclei
Gland
Clustered to 3 cluster by
k-means
2
ABOUT GLANDS
gland can have different sizes and shapes and also nuclei can have very
complex shape.
3
CONTEST CHALLENGE
Challenge: the challenge is to segment glandular structure in colon
histopathology images
4
Typical histopathology image
Segmented gland by
pathologist
DATABASE
The database of GlaS contest consist of 85 train and 60 test Haematoxylin
and Eosin (H&E) stained images, including malignant and benign images.
85 train images
60 test images (Test Part A)
5
PROPOSED GLAND
SEGMENTATION SYSTEM
Lumen extraction
(Label 3 of clustered image)
85 train images
60 test images
(Test Part A)
K-mean
clustering
Structural & Textural
Features extraction
(F_Nuclei)
Combine all Three Features
[F_Lumen, F_Cytoplasm , F_Nuclei]
Structural & Textural
Features extraction
(F_Lumen)
Adjacency Cell nuclei
extraction for each
Cytoplasm
(Label 1 of clustered image)
Training Feature data
(from Train images)
Adjacency Cytoplasm
extraction for each
Lumen
(Label 2 of clustered image)
Structural & Textural
Features extraction
(F_Cytoplasm)
Classification
(SVM, RF)
Testing Feature data
(from Test images)
Overview of proposed gland segmentation system (SUTECH team)
6
gland segmentation
result
PROPOSED GLAND
SEGMENTATION SYSTEM
 Fist step: K-means clustering
Cluster all images to 3 cluster: Lumen, Cytoplasm, Cell nuclei
gland
Lumens
(with remove small objects)
Cytoplasm
(with opening with SE disk 2)
Cell nucleis
(with remove small objects)
7
PROPOSED GLAND
SEGMENTATION SYSTEM
 Second step: Lumen extraction
gland
Lumens
(with remove small objects)
8
PROPOSED GLAND
SEGMENTATION SYSTEM
 Third step: Textural and structural Feature are extracted from each lumen
including :
 Textural: GLCM, CLBP, Gabor, gray level moment, …
 Structural : number of nuclei pixels (label 1) in outer border of lumen,
number of cytoplasm pixels (label 2) in outer border of lumen, area of all
holes in lumen, number of nuclei holes in lumen, area of nuclei holes in
lumen, number of cytoplasm holes in lumen, area of cytoplasm holes in
lumen, number of other lumen holes in lumen, area of other lumen holes in
lumen, Area, Euler Number, Perimeter, Extent, Eccentricity, , …
9
PROPOSED GLAND
SEGMENTATION SYSTEM
 Fourth step: Adjacency Cytoplasm extraction for each Lumen
gland
Lumens
(with remove small objects)
Adjacency Cytoplasm extraction
for each Lumen
(with opening with SE disk 2)
10
PROPOSED GLAND
SEGMENTATION SYSTEM
 Fifth step: Textural and structural Feature are extracted from each
cytoplasm including : (as Lumen Features)
 Textural: GLCM, CLBP, Gabor, gray level moment, …
 Structural
11
PROPOSED GLAND
SEGMENTATION SYSTEM
 Sixth step: Adjacency Cell nuclei extraction for each Cytoplasm
gland
Lumens
(with remove small objects)
Adjacency Cytoplasm extraction
for each Lumen
(with opening with SE disk 2)
Adjacency Cell nuclei extraction
12
for each Cytoplasm
(with remove small objects)
PROPOSED GLAND
SEGMENTATION SYSTEM
 Seventh step: Textural and structural Feature are extracted from each
nuclei ring including :
 Textural: GLCM, CLBP, Gabor, gray level moment, …
 Structural
 Eighth step: Combine three features that extracted from lumens,
corresponded cytoplasms and corresponded nuclei rings.
13
PROPOSED GLAND
SEGMENTATION SYSTEM
 Ninth step: train SVM classifier with kernel RBF by train features and
evaluate the test ones with trained classifier.
2
RBF : K ( xi .x j )  exp(  xi .x j )
 Evaluation metric :
F_measure=
2×Precision×Recall
Precision+Recal
14
SOME OBTAINED RESULTS
15
SOME OBTAINED RESULTS
16
SOME OBTAINED RESULTS
17
SOME OBTAINED RESULTS
18
SOME OBTAINED RESULTS
19
Q &A
Thanks
SUTECH team presentation
by Ramin Nateghi
Q &A
20
Download